Machine learning based intratumor heterogeneity related signature for prognosis and drug sensitivity in breast cancer

Intratumor heterogeneity (ITH) is involved in tumor evolution and drug resistance. Drug sensitivity shows discrepancy in different breast cancer (BRCA) patients due to ITH. The genes mediating ITH in BRCA and their role in predicting prognosis and drug sensitivity is not yet elucidated. An ITH-related signature (IRS) was built by ten methods-based integrative machine learning programs using TCGA, METABRIC and five GEO datasets. Several indicating scores were employed to evaluate the correlation between IRS score and immune microenvironment. The biological role of PINK1 was investigated using CCK-8 assay. The optimal prognostic signature for BRCA cases was the IRS developed using StepCox(both) + Enet(alpha = 0.9) method, which had the highest average C-index of 0.79. IRS acted as a prognostic biomarker and showed good performance in predicting the prognosis of BRCA patients. Lower IRS score indicated high levels of immuno-activated cells, higher TMB score, higher PD1&CTLA4 immunophenoscore, lower ITH score, lower TIDE score and lower tumor escape score in BRCA. The gene set scores correlated with glycolysis, angiogenesis, NOTCH signaling and hypoxia were higher in BRCA with high IRS score. PINK1 knockdown significantly inhibited the proliferation of BRCA cells. Our study developed a novel IRS for BRCA, which could predict the clinical outcome and immunotherapy benefits of BRCA patients.